@Ada321 Phantom works with this lora too

It's in the description: "Example above was generated in about 35 seconds on a 4090 using 4 steps, lcm, 1 cfg, 8 shift."

@funscripter627 Yes, I can see it, but it's example only, as I wanna know recommended steps for this lora.

I'm having better results staying with uni_pc, 6 steps and 9 shift. lcm gets me some weird movement.

Yeah, it seems like Kijai's wrapper itself is spoiling all the fun. Running native WAN results in actually good outputs somehow.

@Ada321 Yeah, just native Comfy WanImageToVideo and that's it. 1 cfg, you can set shift or don't, I can't see any difference. Almost any sampler works.

And by the way I am loading it through power lora loader from rgtree, maybe it somehow matters.

Getting some good results with lcm and native nodes now too. Shift seems to have a way bigger influence than I'm used to (or I just never realized lol) I'm testing with the Phantom base gguf model btw.

Also, don't recommend using rewards lora since it's replacing original faces and bodies somehow.

kindly please share workflow

@kumarkishank959811 https://files.catbox.moe/gufe6r.json

@Ada321 I'm using it with native WAN at 1.0 with no problems. Literally, almost no sampler except dpm-based is causing troubles in generations, either 4, 6 or 8 steps. Simple scheduler is our god and savior.

Also, you could try out my set of nodes, there is a prompt enhancer with an ability to write system prompts and set top_k, top_p, temperature and max tokens: https://github.com/olivv-cs/ComfyUI-FunPack

Wow, I'm getting lost now. So what combination of Wan 2.1 model and what lora should I be using?
Fusion X Wan model with LightX2v lora?

@osakadon Just use the base WAN T2V/I2V model with a self-forcing LoRA, it's better than CauseVid+AccVid combined already.

FusionX while it's good it's already has several speed loras merged in (CauseVid, AccVid, MPS, and Moviigen), so you don’t want to use it with a self-forcing lora, it’s already too diluted with speed loras, imo.

@fronyax Do you have a recommendation workflow to use with FussionX? I'm new to Wan and video generation.

@osakadon If you're okay with not using ComfyUI and you just want to try AI video gens right now without much hassles, maybe try WanGP by DeepBeepMeep (google it). It has most of the best video gen models out there, all in a simple gradio interface. Easiest way to install is via Pinokio app. It's called "Wan 2.1" there.

@Choco7172 I'd ultimately like to get comfortable with ComfyUI, but I'll check your suggestion too.
So if I want to use FussionX I can just find a simple gguf workflow that only needs the video model and no lora and I should be good to go?

@osakadon You mean for ComfyUI or WanGP? If ComfyUI, I'm afraid I'm not the best person to answer that question because I've never used it before :X But WanGP itself has added support for FusionX (just a few days ago). It also support AccVid, CausVid, and most LoRAs (like maybe 99.99% of all LoRAs shared here in CivitAI). All without the need to find or install any workflows, just install Pinokio, install WanGP and it's good to go (it will download the model automatically the 1st time you tried to generate a video using a particular model tho, so don't be surprised if it looks to be not running). WanGP sadly doesn't support GGUF at the moment but most models can be run on as low as 6GB VRAM (the name WanGP refers to "Wan GPU Poor" so yeah, it's aimed for anyone with less than 12GB VRAM). The dev is very active, like he does updates a few times a week or so when there's new models or LoRAs released.

@Choco7172 I tried to install WanGP (without pinocio), but when I ran it, it gave me an error. I tried to install it again. Same error, got frustrated and gave up with it.

@DigitalGarbage could you kindly share me the links for the other loras you are using in the power lora loader? the usual_lora_booster, Wan14B_RealismBoost, DetailEnhancerV1, usual_lora-v3, Wan2.1-Fun-14B-InP-MPS.

Thanks :)

If you use Kijai's version, connect block_swap_args of the WanVideo BlockSwap node to the WanVideo Model Loader node and set the value of blocks_to_swap of the WanVide BlockSwap node to 20-40. Larger value requires smaller VRAM (but it requires enough main memory).

OR WanVideo VRAM Management node to the Model Loder node.

I have 16gb of VRAM too, try virtual_vram_gb to 0.0 in the latest workflow, to my surprise this helped me avoid OOM.

Id say a good place to start with one sampler is: lora strength 1. 8 steps. sampler- euler_ancestral . scheduler- beta.

I get some what ok results with these.

UPDATE: same settings as before, but using sampler DPMPP.2M and SGM_UNIFORM scheduler. Starting to look much better!

can you share your work flow

@kumarkishank959811 i use a slightly self modified version of this runpod template: https://civitai.com/models/1317373/runpod-wan-21-img2video-template-comfyui

Yes with 12g you need to block swap. I have a 4080 and I swap around 20 blocks

Disable TorchCompileModel node if you use native WF
I get stuck as well with it , idk why , both inductor/cudagraph get stuck
Kijai workflow seems to work better

@funscripter627 Im using block swap. Any recommended settings on the block swap? Are you using the default 10?

@blo01 I turned off that node - no difference :(

@Ada321 thanks. I had to switch Browser. Librewolf would not allow the download.

share your fast workflow gguf including loras as well

Can you share the workflow with end frame? much appreciate

Would love to see your workflow!

download a workflow and drag it into comfyui. It will show all missing nodes and the required models. Download the specified model from civitai and place it into required model folder.

espinozaa where to get the workflow?

I'm getting a KSampler Triton not found error even though ComfyUI's python says it's installed. Any ideas?

Is it artifacts? If so, that usually happens at 6 steps. Also try the causvid lora strength somewhere between 0.2-0.5 and you could try a higher CFG like 6 instead of 1?

Clearing VRAM after each generation helped me. Or use the new template which has it intergrated

sorry I have 4gb vram only 🤭

With torch.compile, 832x480@5s only use 6-7gb vram. Fp16 model, the 28.6gb one. I can run 720p@5s with 10gb vram.

With 64gb ram, there will be some offloading to ssd, it only affects first step or so, 96gb is the new bar.

Native nodes.

@shab987 I generate 49 frames 480x720 videos with 4gb vram and 32gb ram via ComfyUI. It's pretty slow of course 🫠

Yes, you need to swap some blocks or lower the resolution and length. There should be a blockswap node near the model loader. I swap around 20 blocks with my 12G vram and 32GB RAM

Also length (frame size) directly multiplies the amount of RAM needed:
See you might be able to run it 61-81 frames, but even with 80GB you can't run it at 400 frames :)

I had the same issue until adding the blockswap node HIGHLY RECOMMEND for no OOM. I set mine to 40. All you gotta do is just add it in between your model loader and lora loader. After doing so, I can now run the 720p q8 model (16gbvram btw) instead of the 480p q4. I can go all the way up 720x1280 w/o OOM with 10 minute generation but i like sticking to semi-low res and upscaling later just for faster gen times. I'm also using other optimizations too

@bhopping thanks ill try what you say!

@funscripter627 Thanks i'll try that.

@yorgashlol yeah i'll keep frames reasonable then 

if the lora isn't trained for 2d that would make sense but the example above is 2d xD

It's CRAZY good for i2v! Thank you so much!
Movement from the first frame, it's actually doing almost exactly what i say in the prompt, fast generation, decent quality.. that's a game changer for me!

2 min which gpu ?

do u have a workflow for use basic wan 2.1 and this new Self-forcing LoRa?

Hi, would you be so kind as to share a full list of what you're using. I'm getting great results from the FusionX model and lora in SwarmUI but Everything i try with Lightx2v is garbage, like just awful quality, noise and artifacts everywhere. If you'd let me know what Base model you're using. quantised? sampler and scheduler, steps, sigma shift, any other settings i might be missing. I thought i hasd a good handle on SwarmUI but Lightx2v just won't play ball with me. I'm on 16gb vram but FusionX is looking stunning on that. Any help hugely appreciated.

@amazingbeauty 4090

@Vyxen808 i don't use workflows, i use SwarmUI :) 

@TheFunk 
I use the wan2.1-i2v-14b-480p-Q4_K_M.gguf model

Sampler UniPC / Scheduler Simple

Steps 5

CFG 1

Sigma Shift i don't even know what that is or if i can set this in SwarmUI :D

The end result quality seems very dependent on the input image quality for me.

how does that prevent video degradation ? Just looking at the names looks more like preventing OOM exceptions

@Sobsob_ If you generate many videos without restarting comfy UI - generated videos start to ignore the starting image and then just become full of green artifacts.

"Clean VRAM Used" and "Purge VRAM" nodes prevent that from happening.
I've seen those nodes used in other workflows.

@dulburis oh ok thanks, i'll use that too then.

@dulburis May I ask where in the workflow you should place these nodes?

complete nonsense, fix your shit... It probably has something to do with using the wanvideowrapper which is terrible. Just use --lowvram --reserve-vram 1.0 as start up options and use native nodes. wanvideowrapper nodes are slow and if it oom it does not clear the vram and you have to restart comfy.

I tried your new work flow but they key take away is that Genning with 24fps gives you much better results but then we get 3 seconds vid instead of 5 seconds. I suspect self-forcing used 24fps distilled.

Gen speed is indeed better but movement is worsen. I don't know if self-forcing's a net gain.

can you share the workflow? i dont know how to do the magic "2 steps at cfg 3 and 3 steps at cfg 1"

@keyblade my current workflow: https://files.catbox.moe/ar35ny.json

example: https://files.catbox.moe/co4oqk.mp4

I'm currently testing NAG instead of CFG, the proper nodes just came out. There's a toggle button to use CFG instead.

I cleaned up the workflow in the json, removed some optional nodes you might not want installed (they need manual tweaking or the newest version doesn't work), one is a faster upscaler using waifu2x and the other is just to force the "before upscale" video to actually output before the upscale, which randomly doesn't happen otherwise. If you want them, the workflow in the example video has them.

@wewewew Thank you very much! I will study it carefully

Could you provide the link?

https://files.catbox.moe/dy4s8g.json

@p1042779030337 Actually that workflow dy4s8g is giving me out of memory, I have two cards installed & replaced few nodes for multigpu workflow, RTX 4080 16 GB & a GTX 1060 6 GB, maybe this workflow needs more than 24 GB or I am doing something wrong here.

Well, it sounds strange. I'm using it on a T4 with 15G. And it's also pretty tweakable. And I'm seeing now that OP put it back on the front page.

@p1042779030337 I made further investigations & learned some new things I never imagined !! , in my case having this combo of two different generations of nvidia cards, is a situation that PyTorch can't deal well with it (the workflow you posted here should work Ok with other scenarios, but people like me, they need some few pre-cautions, like just offload/shift lightweight tasks to the 2nd GPU (like Clip , ClipVIsion), & use the (Vae Decode tiled) node instead of the standard Vae Decode node,,).

Like that the dy4s8g workflow came out responsive & working & reasonably fast ,, so , yeah some workflows are indeed made well to handle the overload & some are just resources hungry without full benefit, & some multigpu setups are not 100% ok with pytorch if the multiple cards are not from close generations/series.

@p1042779030337  Do you know where the git repository for FlowMatchingSigmas node is in this workflow? ComfyUI manager is not detecting it in the missing node search and Google isn't helping either.

https://github.com/BigStationW/flowmatch_scheduler-comfyui

Just read the notes.

i recommend to edit node file and change max shift to 20 or 30

Use fusionX Lora with it. (or if you want full control fusionx lora ingredients and change strange of MPS or other node).

@flo11ok874 interesting i'll try that

Use a 2 step workflow : first part without lightXV for movement, then upscale with LiightXV wanFun controlnet

@Sobsob_  My ability to make workflows is limited, besides i've found on civitai the "FusionX_Ingredients_Workflows" that has great movement and it's quite fast, and no OOM for me using the gguf version, so for now i'll wait and see if this self forcing wfs are actually going anywhere, speed is not everything if the result is not worth it.

@flo11ok874 It gave a bit more movement using Wan2.1_I2V_14B_FusionX_LoRA.safetensors but still rather stiff, however the image colors were very saturated, so not worth it imho.

@skyrimer3d That's way I told you to try FusionX ingredients (workflow has 5 new Loras with links to download) (this 5 Loras was before at FusionX, when You have every ona single node you can change each strengh for better results). Also u can easier try LightX2v lora @ 1.0 or 0.8 + FusionX Lora @ 0.3 or 0.4

I have the same 24 GB. Try using it "wan2.1-i2v-14b-480p-Q8_0.gguf"

@2027rf You are using it on I2V and it works? that's great, I only use WAN with I2V.

Please share, because I keep getting errors with this one.

@osakadon I will create an article with it soon! Follow me to get it

Do you use it with FusionX Wan or just the original Wan 2.1?

@J1B OG WAN 2.1 Q-8

@osakadon Just use standard ComfyUI wan + lora wk with native nodes. Add Lightx2 as Lora and thats it. Sampler LCM, 4 steps, lenght 81

@flo11ok874 yes and no - that sampler from the description is VERY good !!! do add it to your "regular" WF instead ;)

Did you test it for I2V? Does it work?

@Eternal yes it works

Of course! The second sampler node is connected to nothing! I think the uploader has been trolling

@Ada321 yes this one I was using before. This is the better version.

yes cause 4 steps and 1cfg.

Mess with samplers and schedulers. Let's just say I discovered the 'recommenced' ones were ruining the video quality with accelerator LoRAs. So much time wasted on more faulty 'common knowledge'. A colormatch node off your ref image should fix a lot of colour issues (tho not the ones caused when the model is going crazy).